How do you run local AI models on a Mac?

What is Apple Harbor and how does it run local AI models?

Harbor is an unreleased Mac app from Claris, the Apple subsidiary best known for FileMaker. A listing spotted on 10 November 2026 describes it as a way to turn an Apple silicon Mac into a private AI inference server: open-weight models run on the machine, prompts stay on the hardware, and nothing goes to a cloud provider.

The feature that separates it from a normal desktop chat app is memory pooling. Connect two or more Thunderbolt 5 Macs and Harbor treats their unified memory as one resource, so a model too large for a single machine can still load, or the cluster can serve more simultaneous requests. MacRumors reports the listing promised automatic detection of connected Macs with no manual configuration and no command line.

What the listing did not include: a price, a release date, or any statement from Apple. MacRumors says the app is in early access and not publicly available, and that the App Store page has since been removed. The sign-up form asks for a company name, which reads as a business-first launch rather than a consumer download.

What Apple shipped underneath: RDMA over Thunderbolt 5

Harbor is not the first way to cluster Macs. It is the first one that might not need a terminal. Apple's macOS Tahoe 26.2 release notes list RDMA over Thunderbolt as a new feature that enables low-latency communication between Thunderbolt 5 hosts, naming distributed AI inference with MLX as a use case.

Apple's developer note TN3205 adds the constraints. RDMA over Thunderbolt is available starting with macOS 26.2, only on Apple silicon Macs with Thunderbolt 5. It stays off until you reboot into macOS Recovery, open Terminal, run rdma_ctl enable, reboot again, and confirm the interfaces with ibv_devices. Developers get an RDMA Verbs API that works independently of MLX.

That Recovery trip is the gap Harbor is pitching against. A buyer who wants a cluster today follows a technote and a command line; Harbor's listing claims setup in minutes with no manual configuration. Neither route changes the hardware requirement: two Thunderbolt 5 Macs, cabled directly, running macOS 26.2 or later.

What you can run today: LM Studio, Exo, and Ollama

LM Studio is the closest match to Harbor's single-machine promise. Its pricing page lists a $0 tier that runs local LLMs through llama.cpp and MLX, includes offline voice transcription, and allows LM Link on up to five devices. The paid tiers are about hosted inference, not local capability: Bionic+ at $20/month adds US-hosted open-weight models such as Kimi K3, GLM 5.3 and DeepSeek V4 Flash plus discounted bulk tokens, and Pro at $100/month multiplies the usage limits.

Exo covers the clustering half. The project describes itself as connecting all your devices into an AI cluster, with day-0 support for RDMA over Thunderbolt. Its GitHub repository carries an Apache-2.0 license and about 48,300 stars as of 12 November 2026. You pay in setup time: it is a Python project you run yourself, not an App Store download.

Ollama remains the default command-line runner, MIT licensed with roughly 186,700 GitHub stars on the same date. It does not pool memory across machines, so it solves the run-a-model-on-this-Mac problem rather than the run-a-model-bigger-than-this-Mac problem.

Harbor vs LM Studio vs Exo at a glance

Three different jobs, two of which you can start this afternoon. The table separates what each tool does today from what Harbor is reported to do when it ships.

A note on the Harbor column: every value there comes from a listing quoted by MacRumors on 10 November 2026, not from an Apple product page. Treat it as a preview, not a purchase decision.

The honest catch: no price, no date, and a hardware bill

Harbor's most useful feature is also the most expensive one. Pooling memory across Macs means buying a second Mac, Thunderbolt 5 cables, and machines new enough to have Thunderbolt 5 ports. A single Mac with more unified memory is the simpler answer for most buyers, and it avoids the RDMA configuration entirely.

Pricing is unknown. MacRumors notes that Claris sells FileMaker to businesses and that the sign-up page asks for a company name, so a paid or per-seat model is plausible; it could also ship free with macOS. Until Apple publishes something, any chart that puts a number on Harbor is guessing.

Availability is just as open. 9to5Mac expects a launch in the coming weeks or months, and the App Store listing that appeared and vanished is not a ship date.

Who should wait for Harbor, and who should not

• Wait if you already own two Thunderbolt 5 Macs, want the weights on your own hardware, and will not touch a terminal. Harbor is the only option aimed at exactly that buyer.
• Do not wait if one Mac is enough. LM Studio's free tier covers local inference now, and its paid tiers only matter if you also want hosted models.
• Use Exo if you are comfortable with Python and want clustering today under a license you can audit.
• Reconsider the whole idea if your usage is bursty. Cloud API tokens bill by the call; a second Mac bills once, whether or not it is busy.

How to choose in one afternoon

Start with the model, not the app. Check the memory the quantized weights need; if they fit on one Mac, install LM Studio or Ollama and stop there. If they do not, compare a second Mac against a month of cloud API spend for the same workload, and only then look at clustering: Exo today, Harbor when it ships.

The number to write down before you buy anything is the memory requirement of the model at the quantization you actually plan to run. Harbor's pitch is that a cluster absorbs whatever does not fit. Your budget still has to absorb the second machine.

At a glance

Local AI model runners compared, 12 November 2026

FAQ

Is Apple Harbor available to the public?

No. MacRumors reported on 10 November 2026 that Harbor is in early access under Apple's Claris subsidiary, that a brief App Store listing was removed, and that no price or release date has been given.

Can you cluster Macs for AI without Harbor?

Yes. Exo (Apache-2.0) clusters devices today, and Apple's MLX can use RDMA over Thunderbolt on macOS 26.2 or later. You enable RDMA from macOS Recovery with rdma_ctl enable, per Apple's TN3205, so expect a terminal.

Is LM Studio free for local AI models?

Yes. The $0 tier runs local models through llama.cpp and MLX and includes LM Link for up to five devices. Bionic+ at $20/month and Pro at $100/month add US-hosted cloud models and higher usage limits rather than local features.

Related reading

best local LLM tools, LM Studio vs Ollama, self-hosted LLM cost

Sources

• Apple TN3205: RDMA over Thunderbolt
• macOS Tahoe 26.2 release notes
• MacRumors: Apple's Claris testing Harbor
• 9to5Mac: Apple has new Mac app coming for local AI models
• LM Studio pricing
• Exo repository (Apache-2.0)
• Ollama repository (MIT)